The Skill You’re Not Practicing
Writing is one of those skills that degrades quietly. You don’t notice you’ve lost it until you sit down to produce something without assistance and the words won’t come, or they come but in the wrong order, or the structure collapses in the middle and you can’t figure out why.
This is the uncomfortable position that many developers, analysts, and knowledge workers are slowly drifting into. Not because AI writing tools are bad, but because they’re good enough to take over the parts of writing that actually build the skill.
The cognitive work of writing isn’t the typing. It’s the sequencing, the choosing, the moment where you have to decide what you actually mean before you can say it. When a tool does that work for you consistently enough, you stop doing it yourself. And that capacity, like any capacity you stop exercising, atrophies.
What Autocomplete Actually Replaces
It helps to be precise about what these tools are doing, because the surface description (“it finishes your sentences”) undersells the cognitive weight of what’s being offloaded.
In a simplified model of writing, you have at least three distinct layers happening simultaneously. There’s the rhetorical layer: what is the point, why does it matter, how do I structure the argument. There’s the compositional layer: how do I translate that structure into paragraphs, where do transitions go, how do I control emphasis. And there’s the surface layer: word choice, grammar, sentence rhythm.
Early autocomplete tools (think T9 on a phone keyboard, or basic email completion) worked almost entirely at the surface layer. They predicted the next word based on frequency. They were convenient but they didn’t relieve any real cognitive load, because word-level prediction doesn’t touch the hard parts.
Modern LLM-based autocomplete (GitHub Copilot for code, Notion AI, Google’s “Help me write,” the autocomplete in Gmail that now spans multiple sentences) operates across all three layers. It will suggest not just the next word but the next claim, the next paragraph structure, the next rhetorical move. It’s filling in decisions that used to be yours.
This is genuinely useful in the short term. The document gets done. The email goes out. But the question is what you practiced while producing it.
The Spacing Effect and Why Practice Structure Matters
Cognitive science has a well-established concept called the spacing effect: skills and knowledge are retained better when practice is distributed over time rather than massed together. The corollary is that if you consistently offload a skill during practice sessions, you don’t build the retention that those sessions should produce.
Writing a first draft is uncomfortable precisely because it forces you to make decisions under uncertainty. You don’t yet know if the structure will hold. You have to commit to a direction before you can validate it. That discomfort is the productive part. It’s what builds the judgment to write well the next time.
When autocomplete reliably resolves that uncertainty for you, you get fewer uncomfortable decision points per hour of writing time. Each suggestion you accept is a decision you didn’t make. Over hundreds of accepted suggestions, across months of daily use, the number of independent writing decisions you’ve actually made starts to look quite small.
This isn’t a hypothetical concern. Researchers studying GPS navigation have documented a measurable decline in spatial reasoning and navigation ability in people who rely heavily on turn-by-turn directions. The parallel isn’t perfect (writing is more complex than navigation) but the mechanism is the same: when a tool reliably handles a cognitive task, the brain has less incentive to maintain the circuitry for it.
The Second-Order Problem: You Can’t Tell It’s Happening
What makes this particularly tricky is that the feedback loop is broken in a specific way. If you stop lifting weights, you’ll notice within weeks that things feel heavier. If your writing ability is eroding, the output quality may actually stay constant or improve, because you’re still producing decent text with AI assistance. The degradation is invisible in the final artifact.
The signal you’d need to catch the problem is a writing task performed without assistance, and those are becoming rarer. If your workflow has integrated AI suggestions deeply enough, you might go weeks without writing a full paragraph from scratch. The atrophy accumulates silently.
This has a specific professional consequence. The situations where you most need unassisted writing ability tend to be high-stakes ones: a real-time conversation where you’re explaining a complex tradeoff, a performance review you’re drafting in a context where AI assistance feels inappropriate, a technical post-mortem where you have to synthesize what actually happened and why. These are the moments when the skill matters most, and they’re the moments when you’ll notice it’s gone.
The Distinction Between Augmentation and Substitution
The tools themselves are not the problem. A smarter autocomplete is not inherently harmful, and there’s something worth pushing back on in the technophobic version of this argument, which tends to imply that any cognitive assistance is degrading. That’s not right.
A calculator didn’t make mathematicians worse at math. A reference book didn’t make scholars worse at reasoning. The relevant variable isn’t whether you’re using a tool, it’s whether the tool is substituting for practice that would otherwise build a skill, or augmenting a capability you already have.
The calculator saved mathematicians from arithmetic, which isn’t the skill math is actually about. The problem with current writing AI is that it risks substituting for the parts of writing that writing is actually about: forming an argument, deciding what matters, finding the structure in your own thinking.
Augmentation would look like: AI helps me spot where I’ve buried my main point, or flags passive constructions I’m overusing, or tells me a paragraph is unclear without replacing the paragraph. It’s operating as a diagnostician, not a ghostwriter. That’s genuinely useful and doesn’t degrade the underlying skill.
Substitution looks like: I start a paragraph, accept a suggestion for the rest of it, move on. The output is fine but I’ve made no decision about what the paragraph should argue or how it should end.
How to Stay Sharp Without Throwing Out the Tools
If you’ve recognized the pattern in yourself, the fix is not to disable autocomplete. The fix is to be intentional about which cognitive decisions you’re making versus deferring.
A useful heuristic: before accepting a multi-sentence suggestion, state (to yourself) what that suggestion is doing rhetorically. If you can’t articulate whether it’s providing evidence, introducing a counterargument, or transitioning to a new point, you haven’t engaged with it as a writing decision. You’ve just accepted output.
A second heuristic: keep a writing practice that is explicitly unassisted. This doesn’t have to be large. A daily journal entry, a weekly internal document drafted without suggestions, a habit of writing the first paragraph of anything before turning assistance on. The goal is to maintain the decision-making repetitions that build the skill, the same way a programmer who uses IDE autocomplete heavily might still occasionally write code in a bare text editor to stay sharp.
The harder recommendation, but the more important one: learn to distinguish between the discomfort of not knowing what to say (which AI assistance correctly helps with, because that’s a research or thinking problem) and the discomfort of not having yet decided how to say something you do know. The second discomfort is productive. Interrupting it with a suggestion is costing you a rep.
What This Means
AI writing assistance is becoming ambient. It’s in email clients, in document editors, in code review tools, in Slack. The aggregate amount of cognitive writing work being offloaded per knowledge worker per day is substantial and growing.
The tools are not going to get worse, and the incentive to keep them on is real: they save time, they smooth rough drafts, they reduce the friction of producing text. But the cognitive cost is real too, and it’s structured in a way that makes it easy to miss until it matters.
The skill you want to maintain is not typing. It’s decision-making under uncertainty, forming coherent arguments, finding structure in your own thinking before committing it to a page. Those capacities are worth protecting deliberately, because the tools you’re using are very good at quietly taking them over.
